AbstractDriver drowsiness is one of the major causes of on‐road accidents. Abnormal eye behavior, steering wheel activity, and vehicle trajectory during different drowsiness stages were studied in detail to overcome the limitations of single‐sensor approaches. Some measures, such as percentage of eyelid closure, maximum close duration, and percentage of nonsteering were analyzed using analysis of variance (ANOVA) methods. Moreover, a two‐stage data fusion framework was developed for the modeling combination of information from different sources. Fisher's linear discriminant was implied as the feature‐level fusion method, and Dempster‐Shafer evidence theory was introduced in the decision‐level fusion process. The results suggest that the recognition system proposed here provided 90.7% accuracy. The reliability and accuracy of the fusion method were significantly higher than those of single sensors. © 2012 Wiley Periodicals, Inc.